Design

LSD is an acute water-soluble drug. This makes dos­ing easy: take a sin­gle dose of 250μg, dis­solve it in some water (re­frig­er­ated), and then con­sume 1/10th for a dose of 25μg. Some on­line anec­dotes mi­cro­dose daily, while oth­ers mi­cro­dose every 2 or 3 days.

I or­dered 2 tabs ⁠from Vi­ta­Cat on Silk Road⁠:

He shipped them in an en­ve­lope in an air­tight sealed plas­tic sheet; I kept them re­frig­er­ated, un­der­neath a box to block light, and kept sealed until I opened it and used the first one for ⁠a trip⁠. The re­main­ing tab was kept in the re­frig­er­a­tor until a week later. (Given that peo­ple rou­tinely stash tabs in books and other places, I doubt that the week did it much dam­age.) The next morn­ing, I added 25ml of dis­tilled water⁠^⁠11⁠ into an air­tight mason jar, dropped the blot­ter in, shook vig­or­ously, re­frig­er­ated overnight, and re­moved the blot­ter the next af­ter­noon.

Self-blinding a liq­uid is as straight­for­ward as ⁠self-blinding pills⁠: 2 opaque con­tain­ers (one marked; both kept in re­frig­er­a­tor) into which a mil­li­liter drop­per puts a dose of LSD water and a dose of reg­u­lar water.

This sug­gests the fol­low­ing de­sign: a ran­dom­ized dose on day 1, fol­lowed by days 2 and 3 off, then on day 4, drink­ing the sec­ond con­tainer; on day 7, ex­am­ine the con­tain­ers record­ing whether ac­tive/placebo and fi­nally, start­ing over as day 1 of a new pair of 3-day blocks. The power analy­sis (see next sec­tion) in­di­cates >19 ac­tive days are de­sir­able, so 190+μg are needed; this can be di­luted into an evenly di­vis­i­ble amount like 19ml of water and then 1ml ex­tracted each day. 3 days seems to be the max­i­mum that most mi­cro­dos­ing en­thu­si­asts con­sider a dose ac­tive for, so there shouldn’t be is­sues with tol­er­ance or car­ry­over. I ran this de­sign by Dr. Fadi­man, who did not ob­ject. (This de­sign can be seen as a crossover study⁠: re­peated mea­sures⁠ of the same sub­ject serv­ing as their own con­trol. No washout pe­riod to pro­tect against carry-over, though, since I did not ex­pect no­tice­able ef­fects 4 days out from a mi­cro­dose, and my later analy­sis sug­gests there was none.)

Vi­ta­Cat’s Mayan tabs were ad­ver­tised as 250μg, and I di­luted the re­main­ing tab into 20ml of water, so nom­i­nally each 1ml dose would have 12.5μg of LSD, which is at least twice the level (5μg) a num­ber of peo­ple have claimed ben­e­fits from mi­cro­dos­ing, so even if the tab was over­stated by 100μg, the dose should still be enough to pro­duce any ef­fects.

I took one ad­di­tional step: tem­porar­ily sus­pend­ing my ⁠lithium self-experiment⁠. There are many mostly-negative opin­ions on­line about the in­ter­ac­tions of LSD & lithium⁠, al­though most com­menters seem to be talk­ing about ther­a­peu­tic doses of lithium - which are 10-15x larger than the 10mg of lithium oro­tate which I may have been tak­ing if it wasn’t a placebo week. (It later turned out that it was in­deed a placebo week so my pre­cau­tion was un­nec­es­sary.) But there is no point in risk­ing the de­scribed neg­a­tive ef­fects or the weak­en­ing of the trip, so 2 weeks be­fore, the ex­per­i­ment was paused. Ob­vi­ously I can’t prove that this is suf­fi­cient, but I think it was enough: I was tak­ing 10mg oro­tate every other day (on av­er­age) and the anec­dotes which caused me to place it on hold were from peo­ple tak­ing psy­chi­atric doses which are closer to 500mg, and they gen­er­ally seem to think that the ef­fect of lithium didn’t last for months (which is con­sis­tent with the fairly fast me­tab­o­lism of lithium in the body, ~24 hour half-life); as well, I stopped it weeks be­fore the trip, which went fine, which was it­self a week be­fore start­ing mi­cro­dos­ing. So for the lithium to have been a prob­lem, one would have to pos­tu­late that the lev­els of lithium, al­ready 2 or­ders of mag­ni­tude below those claimed to cause prob­lems, did not no­tice­ably in­ter­fere with the trip, but some­how did in­ter­fere with re­sults spread over 6 months af­ter­wards.

The listed ben­e­fits are to mood, pro­duc­tiv­ity, and cre­ativ­ity. We might also want to check other met­rics to see if any LSD ben­e­fit came at a cost. So I will use my usual met­rics plus a new cre­ativ­ity one:

• Zeo sleep data (pri­mar­ily: la­tency, total sleep, num­ber of awak­en­ings, morn­ing feel)

  I ex­pect lit­tle or no ef­fect on these 4 met­rics, or the more ob­scure ones like sleep com­po­si­tion (light/REM/deep per­cent­ages). Two-tailed tests.

• Spaced rep­e­ti­tion⁠ per­for­mance

  Av­er­age grade of cards re­viewed on a given day; I ex­pect no ef­fect or a ben­e­fit (the sim­u­lated an­neal­ing anal­ogy sounds like it might also work for mem­ory). One-tailed.

• mood/pro­duc­tiv­ity self-rating: 1-5

  3 is a nor­mal day, 2 below-average, 4 a very good day, 5 fan­tas­tic etc. The major pre­dic­tion of the mi­cro­dose the­ory is that mood/pro­duc­tiv­ity/cre­ativ­ity will in­crease, so the final analy­sis should ex­ploit our prior and use a one-tailed test that the rat­ings will in­crease (since higher=bet­ter).

• cre­ativ­ity rat­ing: 1-3

  Sim­i­lar. This is not a met­ric I cur­rently track, but would be a new one. One-tailed.

• ac­tive/placebo pre­dic­tion

  A pre­dic­tion recorded at the end of each block on Pre­dic­tion­Book.com⁠; this is not a de­pen­dent vari­able, but a check for whether there is a re­li­ably no­tice­able sub­jec­tive ef­fect.

• Be­fore/after Open­ness to ex­pe­ri­ence⁠ ques­tion­naires

  Open­ness is one of the Big Five⁠ per­son­al­ity traits, roughly linked with cre­ativ­ity, in­ter­est in nov­elty and va­ri­ety; while Con­sci­en­tious­ness⁠ (think self-discipline & hard-work) tends to in­crease with age, Open­ness seems to de­crease. Per­son­al­ity traits, like IQ, are no­to­ri­ously hard to change & pre­dic­tive of many things about a per­son. MacLeanet al2011⁠, a rare RCT⁠ of psilo­cy­bin⁠, found that Open­ness was still in­creased a few per­cent­age points >1 year after dos­ing. (While not that large in mag­ni­tude, MacLean et al 2011_13ya com­pares it with the re­duc­tion in Open­ness over 4 decades and in­crease from suc­cess­ful an­ti­de­pres­sants or sub­stance abuse treat­ment.) The ef­fect seemed to be dri­ven by those re­port­ing a mys­ti­cal ex­pe­ri­ence; de­spite iden­ti­cal start­ing Open­ness, those re­port­ing a non-mystical ex­pe­ri­ence seemed to see theirs fall. The study has ⁠many weak­nesses, but was still the best such study I knew of up to that point. A later RCT on LSD, Carhart-Har­riset al2016⁠, found large ef­fects 2 weeks after 75μg on both Open­ness and op­ti­mism (d = 0.16/0.56 re­spec­tively). Schmid & Liechti2017⁠, on the other hand, saw a smaller in­crease which was not main­tained 12 months later.

  Since to max­i­mize ef­fec­tive­ness, any full trip should be taken be­fore micro-dosing, this sug­gest 3 sam­ples: before-trip, after-trip, after-micro-dosing.

  My first long Big Five sur­vey re­sult (using ⁠Per­son­al­ity Project) in early 2012_12ya put my Open­ness at the ⁠87^th per­centile; the ⁠sec­ond, 2 days after the trip, was iden­ti­cal (87th per­centile). Since I had no mys­tic ex­pe­ri­ence, this is con­sis­tent with MacLeanet al2011’s analy­sis. A ⁠third fol­lowup in March 2013, after the mi­cro­dos­ing ex­per­i­ment, put me at the 93rd per­centile (need­less to say, the con­fi­dence in­ter­vals⁠ for all 3 over­lap due to the per­centile rank­ing being based on rel­a­tively few ques­tions: ~6 for each per­son­al­ity fac­tor).

We don’t need to worry about time-of-day ef­fects. The LSD mi­cro­doses were at sim­i­lar times each day (right after get­ting up). All the de­pen­dent vari­ables were gen­er­ally recorded/done at con­sis­tent times, ex­cept for the spaced rep­e­ti­tion scores since I some­times re­viewed in the af­ter­noon, but I’ve al­ready in­ves­ti­gated time-of-day ef­fects⁠ in my spaced rep­e­ti­tion per­for­mance and they are triv­ially small both in my dataset & in a multi-million-review dataset drawn from thou­sands of Mnemosyne⁠ users.

Limitations

There are 3 major po­ten­tial prob­lems with this ex­per­i­ment: the re­sults may not gen­er­al­ize be­yond me, the tabs may not have con­tained sub­stan­tial lev­els of LSD, and the LSD may have de­graded while being stored.

Power Calculation

The de­scrip­tions are that the ef­fects are strong & no­tice­able, so a large ef­fect size⁠ sug­gests a small ex­per­i­ment; but with 6 met­rics, we will want to cor­rect for mul­ti­ple com­par­isons⁠ (which need will lower the re­quired p-value in our one-tailed tests). We’ll as­sume a large ef­fect size, p = 0.01, and paired⁠/within-subject⁠ ex­per­i­men­tal de­sign⁠^⁠13⁠:

library(pwr)
pwr.t.test(d=0.8,type="paired",power=0.80,alternative="greater",sig.level=0.01)
 
     Paired t test power calculation
 
              n = 18.47818

19 pairs means 19 ac­tive and 19 placebo doses, 38 days total. This does not seem un­duly oner­ous, but note that we’re as­sum­ing LSD has a strong ef­fect since if we cut the ef­fect size in half to d = 0.4 (a medium-small ef­fect) we need 66 pairs or 132 days! (If the ef­fects were over­sold so d = 0.4 was the cor­rect es­ti­mate, and we went with 19 pairs, we would not have an 80% chance of de­tect­ing the ef­fect but rather a 24% chance. Oh well. “We do what we can, be­cause, we must - for the peo­ple who are still alive.”)

How we treat in­di­vid­ual days will mat­ter: do we av­er­age the data for each block of 3 days into a sin­gle data point, or do we treat each dose as re­sult­ing in 3 data points - the ef­fect on the first day, the less­ened ef­fect on the sec­ond day, and the weak­est ef­fect on the third day? (And then the next dose.) The lat­ter seems to be a bet­ter strat­egy, and so with 250μg we get 25 doses of 10μg, each dose is 3 days so as much as 75 days for ac­tive and a sim­i­lar amount for placebo. (75 pairs gives us a sim­i­lar power cal­cu­la­tion all the way down to d = 0.4.)

In ret­ro­spect (after ob­serv­ing the trends which didn’t reach sig­nif­i­cance), since I spec­i­fied 7 met­rics, it prob­a­bly would have been bet­ter to be con­ser­v­a­tive and spec­ify a significance-level of sig.level=(0.05/7) or 0.0071 rather than sig.level=0.05/5 or 0.0100.

VoI

For back­ground on “value of in­for­ma­tion” cal­cu­la­tions, see the ⁠Adder­all cal­cu­la­tion⁠.

With back­ground, de­sign, and power cal­cu­la­tion out of the way, we can cal­cu­late costs & val­ues.

1. Cost of reg­u­lar LSD mi­cro­dos­ing

   A check of list­ings on the Silk Road 1⁠ in­di­cates that 300μg can be bought for ₿5–6 or $61^$45₂₀₁₃; 300μg trans­lates to 30 doses, or 90 days if I fol­low the 3-day dose pat­tern some rec­om­mend; the yearly cost of LSD is then 45 / 90⁄365 = $248^$183₂₀₁₃. Dis­counted at 5% an­nu­ally, the net-present-value/lifetime-cost of switch­ing to LSD would be 183 / ln 1.05 = $5,086^$3,751₂₀₁₃. We can reuse this fig­ure as the po­ten­tial ben­e­fit of LSD mi­cro­doses, since if the ef­fect is pos­i­tive and large enough to no­tice, then it seems worth roughly 50 cents a day.

   There are 3 major cat­e­gories of ad­di­tional costs: the med­ical risk one runs dur­ing the ex­per­i­ment (prin­ci­pally men­tal), the legal risk, and the rep­u­ta­tional risk.

   1. Schiz­o­phre­nia and other is­sues:

      Ear­lier, I looked into re­search bear­ing on the re­la­tion of LSD and Schiz­o­phre­nia⁠, and con­cluded that the ev­i­dence was most con­sis­tent with LSD hav­ing small health risks (that is, the dam­aged or ill sought out all sorts of drugs in­clud­ing LSD, but LSD did not cause the dam­age or at best caused the ill­ness to sur­face ear­lier) with some cor­re­la­tions of psy­chi­atric ben­e­fit from use⁠; this was over the gen­eral LSD pop­u­la­tion in­clud­ing brain-frying large doses, so the risk for mi­cro­doses ought to be much smaller. Much later it oc­curred to me that the rar­ity of using LSD along with sheer small size of LSD ac­tu­ally re­duces the risk of con­sump­tion com­pared to many other prod­ucts, sim­ply be­cause dan­ger­ous doses of many con­t­a­m­i­nants or poi­sons won’t fit: blot­ters re­port­edly max out at ~500μg. Com­pare that with, say, pirac­etam⁠ where peo­ple take <3g of pow­der daily; if the pirac­etam was just 0.3% con­t­a­m­i­nants, that’s as much as 1mg of the stuff - but a blot­ter can’t even hold 1mg of any sub­stance, much less an ac­tive dose of some sub­stance plus ac­ci­den­tal con­t­a­m­i­na­tion. Over­all, I feel com­fort­able round­ing it to zero. To es­ti­mate the pos­si­ble im­pact on schiz­o­phre­nia⁠, I looked into es­ti­mates of hal­lu­cino­gen con­sump­tion and found that es­ti­mates ranged from 7-12% of the en­tire Amer­i­can pop­u­la­tion (~22-38 mil­lion peo­ple) had con­sumed at some point, which over some­thing like a 60-year lifes­pan (teenager to very old adult) im­plies some­thing like 600,000 hal­lu­cino­gen ‘vir­gins’ per year (a fig­ure echoed in one of Fadi­man’s ar­ti­cles).

   2. Prison:

      Search­ing, I found a 1994 fed­eral Sen­tenc­ing Com­mis­sion doc­u­ment⁠ list­ing 83 sen­tences in­volv­ing LSD that year (while search­ing I also found some in­for­ma­tion in­di­cat­ing that 400μg is con­sid­ered 1 dose, so the mi­cro­dose ex­per­i­ment would in­volve <1 dose!); news ar­ti­cles sug­gest many (most?) of those would be for sell­ers and man­u­fac­tur­ers, so a fur­ther analy­sis would re­duce the risk ac­cord­ingly. In any case, even if we mul­ti­ply by 50 for each state, 600000 / (50 × 83) = 145 sug­gest­ing a rough es­ti­mate of 1 in 145 chance of being con­victed. Le­niency could be ex­pected for a first-time non-violent non-trafficking of­fender, so we could set the cost of this out­come at per­haps just $135,585^$100k₂₀₁₃, or an ex­pected loss/cost of $949^$700₂₀₁₃.

   3. So­cial con­se­quences:

      Some rep­u­ta­tional ef­fects are clearly neg­a­tive: a use of LSD would be a big issue in some roles like run­ning for elected of­fice or being CEO of a large com­pany (un­less one is a sec­u­lar saint like Steve Jobs⁠), es­pe­cially since I would ob­vi­ously be un­able to make an ex­cuse like “youth­ful ex­per­i­men­ta­tion” or “a mo­men­tary lapse of judg­ment”, this writeup demon­strat­ing that my use was thor­oughly pre­med­i­tated; how­ever, this does not bother me inas­much as I am un­fit­ted by both tem­pera­ment and tal­ent for such po­si­tions. (Some­thing sim­i­lar could be said about se­cu­rity clear­ances.) In many more con­ser­v­a­tive re­gions, it would be a sub­stan­tial neg­a­tive, but I am al­ready a poor fit for such re­gions and I don’t know whether an LSD use would be very neg­a­tive at the mar­gin, or whether any of my ex­per­i­ments or es­says have al­ready done the dam­age.

      On the pos­i­tive side, how­ever: in some groups or re­gions like Cal­i­for­nia (which may yet lie in my fu­ture), a use of LSD may be a net pos­i­tive as it in­di­cates val­ued qual­i­ties like open-mindedness. The life­time con­sump­tion rate im­plies that some­thing like 12% of the pop­u­la­tion would not judge me harshly for any use. That I have been care­ful and re­spon­si­ble in my ex­per­i­men­ta­tion, and had mul­ti­ple pur­poses in ad­di­tion to mere recre­ation, will lessen the of­fense for many peo­ple (but pos­si­bly would in­crease my guilt in the eyes of oth­ers). There are many fa­mous LSD users like Bill Gates⁠, Richard Feyn­man⁠, Oliver Sacks⁠, ⁠Fran­cis Crick⁠, Kary Mullis⁠ etc, but they seem to have dis­cussed their drug use only after achiev­ing main­stream suc­cess; worse, the fa­mous users all seem to have been as­so­ci­ated with the 1960s or with com­put­ers, sug­gest­ing tol­er­ance for them may be a pass­ing phase and younger peo­ple not so le­nient (but on the other hand, mar­i­juana le­gal­iza­tion⁠ ef­forts are ac­tu­ally suc­ceed­ing be­fore and dur­ing 2012_12ya, sug­gest­ing that Amer­ica may be more tol­er­ant in the fu­ture).

      All con­sid­ered: I am not sure what the rep­u­ta­tional cost would be, nor do I have good ways to es­ti­mate it, and given the in­vis­i­bil­ity of many rep­u­ta­tional costs - how do you know when some­one didn’t offer some­thing to you or how peo­ple are bad-mouthing you? - I may never know the ac­tual cost to me (com­pared to, say, being ar­rested for LSD pos­ses­sion).

2. How much will the ex­per­i­ment cost to run?

   The pro­ce­dure each morn­ing uses up no more than 5 min­utes so (38 × 5) / 60 × 8 = $34^$25₂₀₁₃ dol­lars of time. The sleep, spaced rep­e­ti­tion, and mood/pro­duc­tiv­ity data I al­ready col­lect, and the cre­ativ­ity self-rating is an­other few sec­onds, and so not worth cal­cu­lat­ing. I’m not sure if I have an ap­pro­pri­ate drop­per/sy­ringe avail­able; that and other sup­plies might tack on an­other $14^$10₂₀₁₃. So 45 + 25 + 10 = 80 dol­lars.

3. Pri­ors:

   The ac­cu­racy of the in­for­ma­tion ranges 20-80% de­pend­ing on the ef­fect size; since it’s in­line with my pre­vi­ous knowl­edge of LSD and based on more than a hand­ful of anec­dotes, I would be some­what op­ti­mistic about there being an ef­fect, so a 50% chance the ef­fect ex­ists.

4. Value of In­for­ma­tion

   Power times prior times ben­e­fit minus cost of ex­per­i­men­ta­tion: 0.2 − 0.8 × 0.5 × 3751 − 80 − 700 = −405 to +720. If we cut out the legal risk, the range is +295-1420 and the ex­per­i­ment well worth doing; our es­ti­mate of the legal risk and the ef­fect size de­ter­mine whether we find it worth doing.

Fadiman Comments

After I com­pleted the ex­per­i­ment and pre­pared the data, but be­fore doing any analy­sis (be­sides the pre­dic­tion ac­cu­racy to check the blind­ing), as a form of result-blind re­view­ing⁠, I emailed Fadi­man the fol­low­ing ques­tions & ask­ing per­mis­sion to quote him, and he replied:

1. You’ve seen my plan­ning and method­ol­ogy write up, and I’ve at­tached the cur­rent ver­sion of my full writeup with every­thing I’ve added or changed since (mostly more back­ground, a log of dates, that sort of thing) if you want to reread it. Do you have any ob­jec­tions to how I did it?

Most im­por­tant: I am deeply im­pressed and pleased with your to­tally orig­i­nal and clev­erly de­signed ex­per­i­ment. That you are able to add a double-blind to a one per­son self-study is a de­light in it­self. I am in­cred­i­bly happy to see how many dif­fer­ent ways you went at this. In terms of the gen­eral write up, I hap­pened to be an ar­dent fan of a mix of data, per­sonal re­flec­tion and ob­ser­va­tion. Sci­en­tific writ­ing all too often does every­thing it can to make the sub­ject under scrutiny as dull as pos­si­ble in the writ­ing, as un­in­ter­est­ing as pos­si­ble, and heaven for­bid, to in­clude any gen­uinely real and per­sonal com­men­tary. So I love the way you did it ex­actly. The changes all seem to me ex­pan­sions and im­prove­ments.

2. Have you or any­one else as far as you know, done a long blind/ran­dom­ized/both self-experiment on LSD mi­cro­dos­ing?

Short an­swer. No. Long an­swer: no one. Longest an­swer - the num­ber of peo­ple are doing self-experiments with LSD and other psy­che­delics and even MDMA⁠ is ex­pand­ing. No one is even ap­proach­ing what you’re doing.

3. If the re­sults turn out to be non-statistically-significant or have very small ef­fect sizes, why do you think they might have turned out that way? Which rea­sons would you have most con­fi­dence in as an ex­pla­na­tion of null re­sults?

I do have some guesses as to why non-statistical-significance might occur, and that is putting aside just the ob­vi­ous sta­tis­ti­cal pos­si­bil­ity of using small sam­ple sizes. I will be ac­tu­ally very con­tent to see null re­sults since it will force me and oth­ers to dig more deeply for pos­si­ble rea­sons. So far, you are the first who might pro­duce such re­sults and cer­tainly since you’re the first double-blind, it adds to the value of what you’re doing. As you prob­a­bly know there’s a big con­tro­versy in the reg­u­lar med­ical lit­er­a­ture about how many se­ri­ous drug stud­ies of multibillion-dollar drugs and up with no [statistically-]sig­nif­i­cant dif­fer­ences a few years out from their sales launch and al­most al­ways from university-based stud­ies, not being paid by the phar­ma­ceu­ti­cal com­pany.

In psy­che­delic stud­ies, 100% suc­cess rates are much harder to de­scribe to non-psychedelic au­di­ences used to more con­ven­tional and much lower suc­cess rates about al­most every­thing.

One pos­si­ble rea­son, that only you can an­swer, and you may have done so in the body of the text, but if so I missed it. I’m not sure that you took a cou­ple of non-blind micro-doses be­fore­hand to be sure that you could no­tice a dif­fer­ence with every placebo and pos­i­tive in­ten­tion going with you to no­tice the dif­fer­ence to es­tab­lish a per­sonal base­line of aware­ness of the sub­stance. If that was not done, then it’s not at all clear that the very sim­ple mea­sure­ments used on a daily basis were sen­si­tive enough to pick up days on and days off. The more usual stud­ies from other peo­ple that I’ve seen, (none of them even at­tempt­ing a double-blind or any blind at all,) are look­ing for a set of in­ter­nal changes that they have pre­vi­ously ex­pe­ri­enced. These may be as sub­tle as eat­ing a health­ier lunch, doing a few more reps at a gym, being able to focus on the task an hour longer and so forth. I have a cou­ple of write-ups in my book and when you get to them, you can see that we’re deal­ing with fairly in­tri­cate in­ter­nal mea­sures which you also cover in de­tail now and then in your per­sonal com­men­taries.

That’s my fa­vorite pos­si­ble rea­son. An­other is that, while some peo­ple are very aware on the day of the micro-dose, oth­ers are clear that it is a two-day ex­pe­ri­ence, and a few even longer. If even a tiny residue of the micro-dose is still ac­tive at the time of the next dose, it might make it harder to dis­tin­guish and on from an off day.

How­ever, I’m also aware that also for some peo­ple, the dose is too low.

The re­port in my book that I find most charm­ing, a woman who’s been micro-dosing for about 8 years - ex­cept dur­ing her preg­nancy - uses 20 µg, which for me would be per­cep­tual. So an al­ter­na­tive rea­son, even though the least in­ter­est­ing one, is that your dose was too low. A few peo­ple had felt it was bet­ter to drop down to 7 or 8µg to not get too high. Human vari­abil­ity.

I’m sure I can come up with more com­pli­cated rea­sons, but you asked what my thoughts were be­fore we com­pleted the analy­sis and there you are.

4. Do you ex­pect the re­sults to turn out pos­i­tive? (By pos­i­tive here, I mean statistically-significant after mul­ti­ple cor­rec­tion and d > 0.1.) How much so?

I did ex­pect the re­sults to turn out pos­i­tive be­cause there are real ef­fects that have been re­ported, not only in my lit­tle stud­ies, but for the past few thou­sand years in in­dige­nous so­ci­eties who of course have ex­per­i­mented with micro-doses and every other con­ceiv­able form of use.

As to how [statistically-]sig­nif­i­cant the sta­tis­ti­cal break­out would be - I’ve not re­ally thought much about it. This was your study and your level of sta­tis­ti­cal so­phis­ti­ca­tion clearly ex­ceeds my own. I tend to be some­what sus­pi­cious of any study where the sta­tis­tics are fairly ex­otic, since I know from watch­ing my grad­u­ate stu­dents PhD dis­ser­ta­tions, that when re­ally sim­ple and al­most in­tu­itively ob­vi­ous sta­tis­tics don’t show re­sults, they start to shop around until they find some­thing that gives them a pos­i­tive re­sult even if it’s a sta­tis­tic that no­body else seems to know about. I watch sta­tis­ti­cal con­sul­tants as­sure stu­dents that they will find some­thing that will jus­tify hir­ing them. Is this pure sci­ence as God and Descartes would have it? Not in my book but hey, the real world is never as clean as the data de­rived from it.

5. Are you will­ing to be quoted on any of the above?

Of course, you are free to quote me in any­thing you write, and I’m free to com­plain that I was quoted out of con­text or any­thing else I choose to say to cover my tail. But in this case, you did the work - you asked me to re­view the work - and there­fore my re­view be­comes part of the total pack­age.

I did this for much the same rea­son as I wrote down the exact met­rics and analy­sis in ad­vance: to spec­ify what re­sults are ex­pected, how one will in­ter­pret them, and to elim­i­nate the temp­ta­tion to fudge or mod­ify or spin or self-deceive about any of it.

Analysis

+++ mnemosyne/mnemosyne/example_scripts/export_stats.py 2013-03-10 21:54:04 +0000
@@ -12,7 +12,7 @@
 data_dir = None
 mnemosyne = Mnemosyne(data_dir)
 
-for n in range(-10, 0):
+for n in range(-275, -1):
     start_of_day = mnemosyne.database().start_of_day_n_days_ago(abs(n))

./bin/python ./mnemosyne/example_scripts/export_stats.py \
     | cut -d ' ' -f 2 | sed -e 's/None/NA/' > ~/mnemosyne.csv

lsd <- read.csv("lsd.csv")
lsd$Date <- as.Date(lsd$Date, format="%m/%d/%Y")
# Zeo's CSV export silently omits missing days; look for them to hand-edit empty rows in
# which(!((as.Date("2012-06-09"):as.Date("2012-09-15")) %in% lsd$Date))
mnemo <- read.table("mnemosyne.csv")
lsd$Mnemosyne <- NA
lsd$Mnemosyne <- mnemo$V1
mp <- read.csv("mp.csv")
lsd$MP <- NA
lsd$MP <- mp$MP
creativity <- read.table("creativity.csv")
lsd$Creativity <- NA
lsd$Creativity <- creativity$V1
write.csv(lsd, file="lsd.csv", row.names=FALSE)

let predictions = [(True,  0.45),(False, 0.55),(True,  0.60),(False, 0.40),(True,  0.40),(True,  0.60),
                   (False, 0.40),(True,  0.40),(False, 0.50),(True,  0.40),(False, 0.50),(False, 0.40),
                   (False, 0.60),(True,  0.40),(True,  0.65),(True,  0.50),(False, 0.40),(False, 0.60),
                   (True,  0.60),(True,  0.75)]
logScore ps = sum $ map (\(result,p) -> if result then log p else log (1-p)) ps
 
logScore predictions
→ -13.4685
 
log 0.50 * fromIntegral (length predictions)
→ -13.8629

Graphing Data

Below are the 8 main de­pen­dent vari­ables, de­picted over time & col­ored by whether it fell in a dose 3-day block or con­trol/placebo block. In de­scend­ing order of im­por­tance:

The 5 sleep vari­ables:

lsd <- read.csv("https://gwern.net/doc/personal/2013-gwern-lsdmicrodose.csv")
# filter out baseline
lsd <- lsd[!is.na(lsd$LSD),]
 
l <- lm(cbind(MP, Creativity, Mnemosyne, Total.Z, Time.to.Z, Time.in.Wake, Awakenings, Morning.Feel)
            ~ LSD, data=lsd); summary(l)
# Response MP :
#
# Coefficients:
#             Estimate Std. Error t value Pr(>|t|)
# (Intercept)   3.0339     0.0898   33.79   <2e-16
# LSD          -0.1430     0.1293   -1.11     0.27
#
# Residual standard error: 0.69 on 112 degrees of freedom
#   (13 observations deleted due to missingness)
# Multiple R-squared:  0.0108,    Adjusted R-squared:  0.00197
# F-statistic: 1.22 on 1 and 112 DF,  p-value: 0.271
#
#
# Response Creativity :
#
# Coefficients:
#             Estimate Std. Error t value Pr(>|t|)
# (Intercept)   1.4068     0.0734    19.2   <2e-16
# LSD          -0.1159     0.1056    -1.1     0.28
#
# Residual standard error: 0.564 on 112 degrees of freedom
#   (13 observations deleted due to missingness)
# Multiple R-squared:  0.0106,    Adjusted R-squared:  0.00179
# F-statistic:  1.2 on 1 and 112 DF,  p-value: 0.275
#
#
# Response Mnemosyne :
#
# Coefficients:
#             Estimate Std. Error t value Pr(>|t|)
# (Intercept)  3.82558    0.01625  235.40   <2e-16
# LSD         -0.00958    0.02340   -0.41     0.68
#
# Residual standard error: 0.125 on 112 degrees of freedom
#   (13 observations deleted due to missingness)
# Multiple R-squared:  0.0015,    Adjusted R-squared:  -0.00742
# F-statistic: 0.168 on 1 and 112 DF,  p-value: 0.683
#
#
# Response Total.Z :
#
# Coefficients:
#             Estimate Std. Error t value Pr(>|t|)
# (Intercept)   519.80       6.36   81.67   <2e-16
# LSD            14.28       9.16    1.56     0.12
#
# Residual standard error: 48.9 on 112 degrees of freedom
#   (13 observations deleted due to missingness)
# Multiple R-squared:  0.0212,    Adjusted R-squared:  0.0125
# F-statistic: 2.43 on 1 and 112 DF,  p-value: 0.122
#
#
# Response Time.to.Z :
#
# Coefficients:
#             Estimate Std. Error t value Pr(>|t|)
# (Intercept)    26.58       2.05   12.95   <2e-16
# LSD             2.04       2.95    0.69     0.49
#
# Residual standard error: 15.8 on 112 degrees of freedom
#   (13 observations deleted due to missingness)
# Multiple R-squared:  0.00425,   Adjusted R-squared:  -0.00464
# F-statistic: 0.478 on 1 and 112 DF,  p-value: 0.491
#
#
# Response Time.in.Wake :
#
# Coefficients:
#             Estimate Std. Error t value Pr(>|t|)
# (Intercept)    21.93       1.80   12.17   <2e-16
# LSD             2.78       2.59    1.07     0.29
#
# Residual standard error: 13.8 on 112 degrees of freedom
#   (13 observations deleted due to missingness)
# Multiple R-squared:  0.0101,    Adjusted R-squared:  0.00128
# F-statistic: 1.15 on 1 and 112 DF,  p-value: 0.287
#
#
# Response Awakenings :
#
# Coefficients:
#             Estimate Std. Error t value Pr(>|t|)
# (Intercept)    7.237      0.382   18.96   <2e-16
# LSD            0.635      0.550    1.16     0.25
#
# Residual standard error: 2.93 on 112 degrees of freedom
#   (13 observations deleted due to missingness)
# Multiple R-squared:  0.0118,    Adjusted R-squared:  0.00297
# F-statistic: 1.34 on 1 and 112 DF,  p-value: 0.25
#
#
# Response Morning.Feel :
#
# Coefficients:
#             Estimate Std. Error t value Pr(>|t|)
# (Intercept)    2.593      0.100   25.92   <2e-16
# LSD            0.370      0.144    2.57    0.011
#
# Residual standard error: 0.769 on 112 degrees of freedom
#   (13 observations deleted due to missingness)
# Multiple R-squared:  0.0557,    Adjusted R-squared:  0.0473
# F-statistic: 6.61 on 1 and 112 DF,  p-value: 0.0114
summary(manova(l))
#            Df Pillai approx F num Df den Df Pr(>F)
# LSD         1  0.142     2.17      8    105  0.036
# Residuals 112
 
# MP is the most important metric; what is the effect size (Cohen's d) here?
(mean(lsd[lsd$LSD==1,]$MP) - mean(lsd[lsd$LSD==0,]$MP)) / sd(lsd$MP)
# [1] -0.1782
(mean(lsd[lsd$LSD==1,]$Creativity) - mean(lsd[lsd$LSD==0,]$Creativity)) / sd(lsd$Creativity)
# [1] -0.1921

p.adjust(c(0.27, 0.28, 0.68, 0.12, 0.49, 0.29, 0.25, 0.011), method="BH") < 0.05
# [1] FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE

png(file="~/wiki/doc/nootropic/lsd/gwern-totalz.png", width = 780, height = 680)
qplot(Date, Total.Z, color=LSD, data=lsd)
dev.off()
png(file="~/wiki/doc/nootropic/lsd/gwern-timetoz.png", width = 780, height = 680)
qplot(Date, Time.to.Z, color=LSD, data=lsd)
dev.off()
png(file="~/wiki/doc/nootropic/lsd/gwern-timeinwake.png", width = 780, height = 680)
qplot(Date, Time.in.Wake, color=LSD, data=lsd)
dev.off()
png(file="~/wiki/doc/nootropic/lsd/gwern-awakenings.png", width = 780, height = 680)
qplot(Date, Awakenings, color=LSD, data=lsd)
dev.off()
png(file="~/wiki/doc/nootropic/lsd/gwern-morningfeel.png", width = 780, height = 680)
qplot(Date, Morning.Feel, color=LSD, data=lsd)
dev.off()
png(file="~/wiki/doc/nootropic/lsd/gwern-mnemosyne.png", width = 780, height = 680)
qplot(Date, Mnemosyne, color=LSD, data=lsd)
dev.off()
png(file="~/wiki/doc/nootropic/lsd/gwern-mp.png", width = 780, height = 680)
qplot(Date, MP, color=LSD, data=lsd)
dev.off()
png(file="~/wiki/doc/nootropic/lsd/gwern-creativity.png", width = 780, height = 680)
qplot(Date, Creativity, color=LSD, data=lsd)
dev.off()

Microdose Effect Length

One final de­tail to go with the pre­vi­ous analy­sis is to take a look at whether there were is­sues with 3-day blocks being a bad choice.

I clas­si­fied by hand each day in the rel­e­vant pe­riod by how many days dis­tant it was from the pre­ced­ing LSD mi­cro­dose (that is, the first day of an LSD block is 0, the sec­ond day is 1, the third day is 2, and so on up to 7, and past that I just round down to 7). The idea is that this lets us ask for a lin­ear fit re­lat­ing MP on days which are n dis­tant from a LSD mi­cro­dose and see if there’s any ap­par­ent trend - it may be that we failed to see a statistically-significant re­la­tion­ship be­cause we lumped in all days to­gether.

model1 <- lm(MP ~ DaysSince, data=lsd); summary(model1)
# Residuals:
#     Min      1Q  Median      3Q     Max
# -1.0760 -0.8587  0.0171  0.1413  1.1413
#
# Coefficients:
#             Estimate Std. Error t value Pr(>|t|)
# (Intercept)   2.8587     0.0986    29.0   <2e-16
# DaysSince     0.0310     0.0206     1.5     0.13
#
# Residual standard error: 0.712 on 173 degrees of freedom
#   (99 observations deleted due to missingness)
# Multiple R-squared: 0.0129, Adjusted R-squared: 0.00721
# F-statistic: 2.26 on 1 and 173 DF,  p-value: 0.134
 
plot(jitter(lsd$MP, 0.5) ~ lsd$DaysSince, xlab="Increasing time since a microdose", ylab="Mood/productivity")
abline(model1)

More gen­er­ally, we can re-run the mul­ti­vari­ate re­gres­sion with days-since as an­other pre­dic­tor and see if it adds any­thing:

l1 <- lm(cbind(MP, Creativity, Mnemosyne, Total.Z, Time.to.Z, Time.in.Wake, Awakenings, Morning.Feel)
             ~ LSD + DaysSince, data=lsd); summary(l1)
# Response MP :
#
# Coefficients:
#             Estimate Std. Error t value Pr(>|t|)
# (Intercept)   3.2090     0.2980   10.77   <2e-16
# LSD          -0.2847     0.2639   -1.08     0.28
# DaysSince    -0.0323     0.0524   -0.62     0.54
#
# Residual standard error: 0.692 on 111 degrees of freedom
#   (13 observations deleted due to missingness)
# Multiple R-squared:  0.0142,    Adjusted R-squared:  -0.00358
# F-statistic: 0.798 on 2 and 111 DF,  p-value: 0.453
#
#
# Response Creativity :
#
# Coefficients:
#             Estimate Std. Error t value Pr(>|t|)
# (Intercept)   1.2435     0.2434    5.11  1.4e-06
# LSD           0.0162     0.2155    0.08     0.94
# DaysSince     0.0301     0.0428    0.70     0.48
#
# Residual standard error: 0.565 on 111 degrees of freedom
#   (13 observations deleted due to missingness)
# Multiple R-squared:  0.015, Adjusted R-squared:  -0.00273
# F-statistic: 0.846 on 2 and 111 DF,  p-value: 0.432
#
#
# Response Mnemosyne :
#
# Coefficients:
#             Estimate Std. Error t value Pr(>|t|)
# (Intercept)  3.80753    0.05400   70.51   <2e-16
# LSD          0.00502    0.04781    0.10     0.92
# DaysSince    0.00333    0.00949    0.35     0.73
#
# Residual standard error: 0.125 on 111 degrees of freedom
#   (13 observations deleted due to missingness)
# Multiple R-squared:  0.0026,    Adjusted R-squared:  -0.0154
# F-statistic: 0.145 on 2 and 111 DF,  p-value: 0.865
#
#
# Response Total.Z :
#
# Coefficients:
#             Estimate Std. Error t value Pr(>|t|)
# (Intercept)   545.47      21.01   25.97   <2e-16
# LSD            -6.49      18.60   -0.35     0.73
# DaysSince      -4.73       3.69   -1.28     0.20
#
# Residual standard error: 48.7 on 111 degrees of freedom
#   (13 observations deleted due to missingness)
# Multiple R-squared:  0.0355,    Adjusted R-squared:  0.0181
# F-statistic: 2.04 on 2 and 111 DF,  p-value: 0.135
#
#
# Response Time.to.Z :
#
# Coefficients:
#             Estimate Std. Error t value Pr(>|t|)
# (Intercept)    32.69       6.80    4.81  4.8e-06
# LSD            -2.91       6.02   -0.48     0.63
# DaysSince      -1.13       1.19   -0.94     0.35
#
# Residual standard error: 15.8 on 111 degrees of freedom
#   (13 observations deleted due to missingness)
# Multiple R-squared:  0.0122,    Adjusted R-squared:  -0.00562
# F-statistic: 0.684 on 2 and 111 DF,  p-value: 0.506
#
#
# Response Time.in.Wake :
#
# Coefficients:
#             Estimate Std. Error t value Pr(>|t|)
# (Intercept)   26.155      5.978    4.38  2.7e-05
# LSD           -0.639      5.292   -0.12     0.90
# DaysSince     -0.779      1.051   -0.74     0.46
#
# Residual standard error: 13.9 on 111 degrees of freedom
#   (13 observations deleted due to missingness)
# Multiple R-squared:  0.015, Adjusted R-squared:  -0.00275
# F-statistic: 0.845 on 2 and 111 DF,  p-value: 0.432
#
#
# Response Awakenings :
#
# Coefficients:
#             Estimate Std. Error t value Pr(>|t|)
# (Intercept)    9.374      1.251    7.49  1.8e-11
# LSD           -1.093      1.108   -0.99    0.326
# DaysSince     -0.394      0.220   -1.79    0.076
#
# Residual standard error: 2.9 on 111 degrees of freedom
#   (13 observations deleted due to missingness)
# Multiple R-squared:  0.0396,    Adjusted R-squared:  0.0223
# F-statistic: 2.29 on 2 and 111 DF,  p-value: 0.106
#
#
# Response Morning.Feel :
#
# Coefficients:
#             Estimate Std. Error t value Pr(>|t|)
# (Intercept)    2.181      0.330    6.61  1.4e-09
# LSD            0.704      0.292    2.41    0.018
# DaysSince      0.076      0.058    1.31    0.193
#
# Residual standard error: 0.766 on 111 degrees of freedom
#   (13 observations deleted due to missingness)
# Multiple R-squared:  0.0701,    Adjusted R-squared:  0.0533
# F-statistic: 4.18 on 2 and 111 DF,  p-value: 0.0177
summary(manova(l1))
#            Df Pillai approx F num Df den Df Pr(>F)
# LSD         1  0.142     2.15      8    104  0.038
# DaysSince   1  0.076     1.07      8    104  0.390
# Residuals 111

None of the es­ti­mates are statistically-significant, but the MANOVA sug­gests that there’s mod­est ev­i­dence that the amount of time since a dose mat­ters. To sum­ma­rize the re­sults for using DaysSince as well in a neat lit­tle table:

Vari­able	Ef­fect	p-value	Co­ef­fi­cient’s sign is…
MP	-0.03	0.54	worse
Creativity	0.03	0.48	bet­ter
Mnemosyne	0.00	0.73
Total.Z	-4.73	0.20	worse
Time.to.Z	-1.13	0.35	bet­ter
Time.in.Wake	-0.78	0.46	bet­ter
Awakenings	-0.39	0.07	bet­ter
Morning.Feel	0.08	0.19	bet­ter

Do we get a bet­ter fit if we omit days “too far” from the mi­cro­dose, rea­son­ing that the ef­fects would’ve stopped? Look­ing just at MP to keep things sim­ple:

for (i in 7:2) { print(summary(lm(MP ~ DaysSince, data=lsd[!(lsd$DaysSince>=i),]))); }

days re­moved	cor­re­la­tion	p
	+0.03100	0.13
7th	+0.00418	0.90
6,7	-0.00514	0.89
5,6,7	-0.05220	0.32
4,5,6,7	-0.06680	0.35
3,4,5,6,7	-0.17400	0.11
2,3,4,5,6,7	+0.00000	1.00

Ob­vi­ously the p-values are mean­ing­less in this ap­pli­ca­tion; more im­por­tantly, the re­ver­sal of co­ef­fi­cient sign and chang­ing co­ef­fi­cient size & statistical-significance don’t sug­gest any­thing in par­tic­u­lar to me about dose pe­riod.